Interpretable machine learning for gallstone risk prediction: an ensemble stacking approach with SHAP analysis

Area of Science:

• Biomedical Engineering
• Gastroenterology
• Machine Learning in Healthcare

Background:

• Gallstone disease presents a significant gastrointestinal health burden.
• Accessible, low-cost screening methods for gallstones are currently limited.
• Early detection of gallstones can improve patient outcomes and reduce healthcare costs.

Purpose of the Study:

• To develop and validate an interpretable ensemble stacking model for noninvasive gallstone prediction.
• To utilize bioimpedance and laboratory data for enhanced diagnostic accuracy.
• To identify key clinical predictors contributing to gallstone formation.

Main Methods:

• An ensemble stacking model integrating Bayesian-optimized Support Vector Machine (SVM), logistic regression, and decision trees with a deep neural network meta-learner was developed.
• Data from 319 individuals (157 patients, 162 controls) with 38 features, including bioimpedance and laboratory values, were used.
• Ten repeated nested cross-validation loops were employed to evaluate model performance.

Main Results:

• The developed model achieved 82.3% accuracy and 85.7% Area Under the Curve (AUC), outperforming classical stacking and eight individual algorithms.
• Shapley Additive exPlanations (SHAP) analysis identified C-reactive protein (CRP), liver enzymes, and HDL cholesterol as significant contributors.
• Multicollinearity issues were noted for vitamin D and diabetes-related features, impacting directional stability.

Conclusions:

• The interpretable ensemble stacking model demonstrates high accuracy for noninvasive gallstone prediction using readily available data.
• Key biochemical markers like CRP, liver enzymes, and HDL cholesterol are crucial for gallstone risk assessment.
• Further prospective, multicenter validation is required prior to clinical implementation.